readahead: split ondemand readahead interface into two functions
[linux-2.6.git] / fs / splice.c
1 /*
2  * "splice": joining two ropes together by interweaving their strands.
3  *
4  * This is the "extended pipe" functionality, where a pipe is used as
5  * an arbitrary in-memory buffer. Think of a pipe as a small kernel
6  * buffer that you can use to transfer data from one end to the other.
7  *
8  * The traditional unix read/write is extended with a "splice()" operation
9  * that transfers data buffers to or from a pipe buffer.
10  *
11  * Named by Larry McVoy, original implementation from Linus, extended by
12  * Jens to support splicing to files, network, direct splicing, etc and
13  * fixing lots of bugs.
14  *
15  * Copyright (C) 2005-2006 Jens Axboe <axboe@kernel.dk>
16  * Copyright (C) 2005-2006 Linus Torvalds <torvalds@osdl.org>
17  * Copyright (C) 2006 Ingo Molnar <mingo@elte.hu>
18  *
19  */
20 #include <linux/fs.h>
21 #include <linux/file.h>
22 #include <linux/pagemap.h>
23 #include <linux/splice.h>
24 #include <linux/mm_inline.h>
25 #include <linux/swap.h>
26 #include <linux/writeback.h>
27 #include <linux/buffer_head.h>
28 #include <linux/module.h>
29 #include <linux/syscalls.h>
30 #include <linux/uio.h>
31 #include <linux/security.h>
32
33 /*
34  * Attempt to steal a page from a pipe buffer. This should perhaps go into
35  * a vm helper function, it's already simplified quite a bit by the
36  * addition of remove_mapping(). If success is returned, the caller may
37  * attempt to reuse this page for another destination.
38  */
39 static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
40                                      struct pipe_buffer *buf)
41 {
42         struct page *page = buf->page;
43         struct address_space *mapping;
44
45         lock_page(page);
46
47         mapping = page_mapping(page);
48         if (mapping) {
49                 WARN_ON(!PageUptodate(page));
50
51                 /*
52                  * At least for ext2 with nobh option, we need to wait on
53                  * writeback completing on this page, since we'll remove it
54                  * from the pagecache.  Otherwise truncate wont wait on the
55                  * page, allowing the disk blocks to be reused by someone else
56                  * before we actually wrote our data to them. fs corruption
57                  * ensues.
58                  */
59                 wait_on_page_writeback(page);
60
61                 if (PagePrivate(page))
62                         try_to_release_page(page, GFP_KERNEL);
63
64                 /*
65                  * If we succeeded in removing the mapping, set LRU flag
66                  * and return good.
67                  */
68                 if (remove_mapping(mapping, page)) {
69                         buf->flags |= PIPE_BUF_FLAG_LRU;
70                         return 0;
71                 }
72         }
73
74         /*
75          * Raced with truncate or failed to remove page from current
76          * address space, unlock and return failure.
77          */
78         unlock_page(page);
79         return 1;
80 }
81
82 static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
83                                         struct pipe_buffer *buf)
84 {
85         page_cache_release(buf->page);
86         buf->flags &= ~PIPE_BUF_FLAG_LRU;
87 }
88
89 /*
90  * Check whether the contents of buf is OK to access. Since the content
91  * is a page cache page, IO may be in flight.
92  */
93 static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
94                                        struct pipe_buffer *buf)
95 {
96         struct page *page = buf->page;
97         int err;
98
99         if (!PageUptodate(page)) {
100                 lock_page(page);
101
102                 /*
103                  * Page got truncated/unhashed. This will cause a 0-byte
104                  * splice, if this is the first page.
105                  */
106                 if (!page->mapping) {
107                         err = -ENODATA;
108                         goto error;
109                 }
110
111                 /*
112                  * Uh oh, read-error from disk.
113                  */
114                 if (!PageUptodate(page)) {
115                         err = -EIO;
116                         goto error;
117                 }
118
119                 /*
120                  * Page is ok afterall, we are done.
121                  */
122                 unlock_page(page);
123         }
124
125         return 0;
126 error:
127         unlock_page(page);
128         return err;
129 }
130
131 static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
132         .can_merge = 0,
133         .map = generic_pipe_buf_map,
134         .unmap = generic_pipe_buf_unmap,
135         .confirm = page_cache_pipe_buf_confirm,
136         .release = page_cache_pipe_buf_release,
137         .steal = page_cache_pipe_buf_steal,
138         .get = generic_pipe_buf_get,
139 };
140
141 static int user_page_pipe_buf_steal(struct pipe_inode_info *pipe,
142                                     struct pipe_buffer *buf)
143 {
144         if (!(buf->flags & PIPE_BUF_FLAG_GIFT))
145                 return 1;
146
147         buf->flags |= PIPE_BUF_FLAG_LRU;
148         return generic_pipe_buf_steal(pipe, buf);
149 }
150
151 static const struct pipe_buf_operations user_page_pipe_buf_ops = {
152         .can_merge = 0,
153         .map = generic_pipe_buf_map,
154         .unmap = generic_pipe_buf_unmap,
155         .confirm = generic_pipe_buf_confirm,
156         .release = page_cache_pipe_buf_release,
157         .steal = user_page_pipe_buf_steal,
158         .get = generic_pipe_buf_get,
159 };
160
161 /**
162  * splice_to_pipe - fill passed data into a pipe
163  * @pipe:       pipe to fill
164  * @spd:        data to fill
165  *
166  * Description:
167  *    @spd contains a map of pages and len/offset tupples, a long with
168  *    the struct pipe_buf_operations associated with these pages. This
169  *    function will link that data to the pipe.
170  *
171  */
172 ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
173                        struct splice_pipe_desc *spd)
174 {
175         unsigned int spd_pages = spd->nr_pages;
176         int ret, do_wakeup, page_nr;
177
178         ret = 0;
179         do_wakeup = 0;
180         page_nr = 0;
181
182         if (pipe->inode)
183                 mutex_lock(&pipe->inode->i_mutex);
184
185         for (;;) {
186                 if (!pipe->readers) {
187                         send_sig(SIGPIPE, current, 0);
188                         if (!ret)
189                                 ret = -EPIPE;
190                         break;
191                 }
192
193                 if (pipe->nrbufs < PIPE_BUFFERS) {
194                         int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
195                         struct pipe_buffer *buf = pipe->bufs + newbuf;
196
197                         buf->page = spd->pages[page_nr];
198                         buf->offset = spd->partial[page_nr].offset;
199                         buf->len = spd->partial[page_nr].len;
200                         buf->private = spd->partial[page_nr].private;
201                         buf->ops = spd->ops;
202                         if (spd->flags & SPLICE_F_GIFT)
203                                 buf->flags |= PIPE_BUF_FLAG_GIFT;
204
205                         pipe->nrbufs++;
206                         page_nr++;
207                         ret += buf->len;
208
209                         if (pipe->inode)
210                                 do_wakeup = 1;
211
212                         if (!--spd->nr_pages)
213                                 break;
214                         if (pipe->nrbufs < PIPE_BUFFERS)
215                                 continue;
216
217                         break;
218                 }
219
220                 if (spd->flags & SPLICE_F_NONBLOCK) {
221                         if (!ret)
222                                 ret = -EAGAIN;
223                         break;
224                 }
225
226                 if (signal_pending(current)) {
227                         if (!ret)
228                                 ret = -ERESTARTSYS;
229                         break;
230                 }
231
232                 if (do_wakeup) {
233                         smp_mb();
234                         if (waitqueue_active(&pipe->wait))
235                                 wake_up_interruptible_sync(&pipe->wait);
236                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
237                         do_wakeup = 0;
238                 }
239
240                 pipe->waiting_writers++;
241                 pipe_wait(pipe);
242                 pipe->waiting_writers--;
243         }
244
245         if (pipe->inode) {
246                 mutex_unlock(&pipe->inode->i_mutex);
247
248                 if (do_wakeup) {
249                         smp_mb();
250                         if (waitqueue_active(&pipe->wait))
251                                 wake_up_interruptible(&pipe->wait);
252                         kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
253                 }
254         }
255
256         while (page_nr < spd_pages)
257                 page_cache_release(spd->pages[page_nr++]);
258
259         return ret;
260 }
261
262 static int
263 __generic_file_splice_read(struct file *in, loff_t *ppos,
264                            struct pipe_inode_info *pipe, size_t len,
265                            unsigned int flags)
266 {
267         struct address_space *mapping = in->f_mapping;
268         unsigned int loff, nr_pages, req_pages;
269         struct page *pages[PIPE_BUFFERS];
270         struct partial_page partial[PIPE_BUFFERS];
271         struct page *page;
272         pgoff_t index, end_index;
273         loff_t isize;
274         int error, page_nr;
275         struct splice_pipe_desc spd = {
276                 .pages = pages,
277                 .partial = partial,
278                 .flags = flags,
279                 .ops = &page_cache_pipe_buf_ops,
280         };
281
282         index = *ppos >> PAGE_CACHE_SHIFT;
283         loff = *ppos & ~PAGE_CACHE_MASK;
284         req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
285         nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
286
287         /*
288          * Lookup the (hopefully) full range of pages we need.
289          */
290         spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
291         index += spd.nr_pages;
292
293         /*
294          * If find_get_pages_contig() returned fewer pages than we needed,
295          * readahead/allocate the rest and fill in the holes.
296          */
297         if (spd.nr_pages < nr_pages)
298                 page_cache_sync_readahead(mapping, &in->f_ra, in,
299                                 index, req_pages - spd.nr_pages);
300
301         error = 0;
302         while (spd.nr_pages < nr_pages) {
303                 /*
304                  * Page could be there, find_get_pages_contig() breaks on
305                  * the first hole.
306                  */
307                 page = find_get_page(mapping, index);
308                 if (!page) {
309                         /*
310                          * page didn't exist, allocate one.
311                          */
312                         page = page_cache_alloc_cold(mapping);
313                         if (!page)
314                                 break;
315
316                         error = add_to_page_cache_lru(page, mapping, index,
317                                               GFP_KERNEL);
318                         if (unlikely(error)) {
319                                 page_cache_release(page);
320                                 if (error == -EEXIST)
321                                         continue;
322                                 break;
323                         }
324                         /*
325                          * add_to_page_cache() locks the page, unlock it
326                          * to avoid convoluting the logic below even more.
327                          */
328                         unlock_page(page);
329                 }
330
331                 pages[spd.nr_pages++] = page;
332                 index++;
333         }
334
335         /*
336          * Now loop over the map and see if we need to start IO on any
337          * pages, fill in the partial map, etc.
338          */
339         index = *ppos >> PAGE_CACHE_SHIFT;
340         nr_pages = spd.nr_pages;
341         spd.nr_pages = 0;
342         for (page_nr = 0; page_nr < nr_pages; page_nr++) {
343                 unsigned int this_len;
344
345                 if (!len)
346                         break;
347
348                 /*
349                  * this_len is the max we'll use from this page
350                  */
351                 this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
352                 page = pages[page_nr];
353
354                 if (PageReadahead(page))
355                         page_cache_async_readahead(mapping, &in->f_ra, in,
356                                         page, index, req_pages - page_nr);
357
358                 /*
359                  * If the page isn't uptodate, we may need to start io on it
360                  */
361                 if (!PageUptodate(page)) {
362                         /*
363                          * If in nonblock mode then dont block on waiting
364                          * for an in-flight io page
365                          */
366                         if (flags & SPLICE_F_NONBLOCK) {
367                                 if (TestSetPageLocked(page))
368                                         break;
369                         } else
370                                 lock_page(page);
371
372                         /*
373                          * page was truncated, stop here. if this isn't the
374                          * first page, we'll just complete what we already
375                          * added
376                          */
377                         if (!page->mapping) {
378                                 unlock_page(page);
379                                 break;
380                         }
381                         /*
382                          * page was already under io and is now done, great
383                          */
384                         if (PageUptodate(page)) {
385                                 unlock_page(page);
386                                 goto fill_it;
387                         }
388
389                         /*
390                          * need to read in the page
391                          */
392                         error = mapping->a_ops->readpage(in, page);
393                         if (unlikely(error)) {
394                                 /*
395                                  * We really should re-lookup the page here,
396                                  * but it complicates things a lot. Instead
397                                  * lets just do what we already stored, and
398                                  * we'll get it the next time we are called.
399                                  */
400                                 if (error == AOP_TRUNCATED_PAGE)
401                                         error = 0;
402
403                                 break;
404                         }
405                 }
406 fill_it:
407                 /*
408                  * i_size must be checked after PageUptodate.
409                  */
410                 isize = i_size_read(mapping->host);
411                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
412                 if (unlikely(!isize || index > end_index))
413                         break;
414
415                 /*
416                  * if this is the last page, see if we need to shrink
417                  * the length and stop
418                  */
419                 if (end_index == index) {
420                         unsigned int plen;
421
422                         /*
423                          * max good bytes in this page
424                          */
425                         plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
426                         if (plen <= loff)
427                                 break;
428
429                         /*
430                          * force quit after adding this page
431                          */
432                         this_len = min(this_len, plen - loff);
433                         len = this_len;
434                 }
435
436                 partial[page_nr].offset = loff;
437                 partial[page_nr].len = this_len;
438                 len -= this_len;
439                 loff = 0;
440                 spd.nr_pages++;
441                 index++;
442         }
443
444         /*
445          * Release any pages at the end, if we quit early. 'page_nr' is how far
446          * we got, 'nr_pages' is how many pages are in the map.
447          */
448         while (page_nr < nr_pages)
449                 page_cache_release(pages[page_nr++]);
450         in->f_ra.prev_index = index;
451
452         if (spd.nr_pages)
453                 return splice_to_pipe(pipe, &spd);
454
455         return error;
456 }
457
458 /**
459  * generic_file_splice_read - splice data from file to a pipe
460  * @in:         file to splice from
461  * @ppos:       position in @in
462  * @pipe:       pipe to splice to
463  * @len:        number of bytes to splice
464  * @flags:      splice modifier flags
465  *
466  * Description:
467  *    Will read pages from given file and fill them into a pipe. Can be
468  *    used as long as the address_space operations for the source implements
469  *    a readpage() hook.
470  *
471  */
472 ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
473                                  struct pipe_inode_info *pipe, size_t len,
474                                  unsigned int flags)
475 {
476         ssize_t spliced;
477         int ret;
478         loff_t isize, left;
479
480         isize = i_size_read(in->f_mapping->host);
481         if (unlikely(*ppos >= isize))
482                 return 0;
483
484         left = isize - *ppos;
485         if (unlikely(left < len))
486                 len = left;
487
488         ret = 0;
489         spliced = 0;
490         while (len && !spliced) {
491                 ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
492
493                 if (ret < 0)
494                         break;
495                 else if (!ret) {
496                         if (spliced)
497                                 break;
498                         if (flags & SPLICE_F_NONBLOCK) {
499                                 ret = -EAGAIN;
500                                 break;
501                         }
502                 }
503
504                 *ppos += ret;
505                 len -= ret;
506                 spliced += ret;
507         }
508
509         if (spliced)
510                 return spliced;
511
512         return ret;
513 }
514
515 EXPORT_SYMBOL(generic_file_splice_read);
516
517 /*
518  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
519  * using sendpage(). Return the number of bytes sent.
520  */
521 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
522                             struct pipe_buffer *buf, struct splice_desc *sd)
523 {
524         struct file *file = sd->u.file;
525         loff_t pos = sd->pos;
526         int ret, more;
527
528         ret = buf->ops->confirm(pipe, buf);
529         if (!ret) {
530                 more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
531
532                 ret = file->f_op->sendpage(file, buf->page, buf->offset,
533                                            sd->len, &pos, more);
534         }
535
536         return ret;
537 }
538
539 /*
540  * This is a little more tricky than the file -> pipe splicing. There are
541  * basically three cases:
542  *
543  *      - Destination page already exists in the address space and there
544  *        are users of it. For that case we have no other option that
545  *        copying the data. Tough luck.
546  *      - Destination page already exists in the address space, but there
547  *        are no users of it. Make sure it's uptodate, then drop it. Fall
548  *        through to last case.
549  *      - Destination page does not exist, we can add the pipe page to
550  *        the page cache and avoid the copy.
551  *
552  * If asked to move pages to the output file (SPLICE_F_MOVE is set in
553  * sd->flags), we attempt to migrate pages from the pipe to the output
554  * file address space page cache. This is possible if no one else has
555  * the pipe page referenced outside of the pipe and page cache. If
556  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
557  * a new page in the output file page cache and fill/dirty that.
558  */
559 static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
560                         struct splice_desc *sd)
561 {
562         struct file *file = sd->u.file;
563         struct address_space *mapping = file->f_mapping;
564         unsigned int offset, this_len;
565         struct page *page;
566         pgoff_t index;
567         int ret;
568
569         /*
570          * make sure the data in this buffer is uptodate
571          */
572         ret = buf->ops->confirm(pipe, buf);
573         if (unlikely(ret))
574                 return ret;
575
576         index = sd->pos >> PAGE_CACHE_SHIFT;
577         offset = sd->pos & ~PAGE_CACHE_MASK;
578
579         this_len = sd->len;
580         if (this_len + offset > PAGE_CACHE_SIZE)
581                 this_len = PAGE_CACHE_SIZE - offset;
582
583 find_page:
584         page = find_lock_page(mapping, index);
585         if (!page) {
586                 ret = -ENOMEM;
587                 page = page_cache_alloc_cold(mapping);
588                 if (unlikely(!page))
589                         goto out_ret;
590
591                 /*
592                  * This will also lock the page
593                  */
594                 ret = add_to_page_cache_lru(page, mapping, index,
595                                             GFP_KERNEL);
596                 if (unlikely(ret))
597                         goto out;
598         }
599
600         ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
601         if (unlikely(ret)) {
602                 loff_t isize = i_size_read(mapping->host);
603
604                 if (ret != AOP_TRUNCATED_PAGE)
605                         unlock_page(page);
606                 page_cache_release(page);
607                 if (ret == AOP_TRUNCATED_PAGE)
608                         goto find_page;
609
610                 /*
611                  * prepare_write() may have instantiated a few blocks
612                  * outside i_size.  Trim these off again.
613                  */
614                 if (sd->pos + this_len > isize)
615                         vmtruncate(mapping->host, isize);
616
617                 goto out_ret;
618         }
619
620         if (buf->page != page) {
621                 /*
622                  * Careful, ->map() uses KM_USER0!
623                  */
624                 char *src = buf->ops->map(pipe, buf, 1);
625                 char *dst = kmap_atomic(page, KM_USER1);
626
627                 memcpy(dst + offset, src + buf->offset, this_len);
628                 flush_dcache_page(page);
629                 kunmap_atomic(dst, KM_USER1);
630                 buf->ops->unmap(pipe, buf, src);
631         }
632
633         ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
634         if (ret) {
635                 if (ret == AOP_TRUNCATED_PAGE) {
636                         page_cache_release(page);
637                         goto find_page;
638                 }
639                 if (ret < 0)
640                         goto out;
641                 /*
642                  * Partial write has happened, so 'ret' already initialized by
643                  * number of bytes written, Where is nothing we have to do here.
644                  */
645         } else
646                 ret = this_len;
647         /*
648          * Return the number of bytes written and mark page as
649          * accessed, we are now done!
650          */
651         mark_page_accessed(page);
652 out:
653         page_cache_release(page);
654         unlock_page(page);
655 out_ret:
656         return ret;
657 }
658
659 /**
660  * __splice_from_pipe - splice data from a pipe to given actor
661  * @pipe:       pipe to splice from
662  * @sd:         information to @actor
663  * @actor:      handler that splices the data
664  *
665  * Description:
666  *    This function does little more than loop over the pipe and call
667  *    @actor to do the actual moving of a single struct pipe_buffer to
668  *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
669  *    pipe_to_user.
670  *
671  */
672 ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
673                            splice_actor *actor)
674 {
675         int ret, do_wakeup, err;
676
677         ret = 0;
678         do_wakeup = 0;
679
680         for (;;) {
681                 if (pipe->nrbufs) {
682                         struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
683                         const struct pipe_buf_operations *ops = buf->ops;
684
685                         sd->len = buf->len;
686                         if (sd->len > sd->total_len)
687                                 sd->len = sd->total_len;
688
689                         err = actor(pipe, buf, sd);
690                         if (err <= 0) {
691                                 if (!ret && err != -ENODATA)
692                                         ret = err;
693
694                                 break;
695                         }
696
697                         ret += err;
698                         buf->offset += err;
699                         buf->len -= err;
700
701                         sd->len -= err;
702                         sd->pos += err;
703                         sd->total_len -= err;
704                         if (sd->len)
705                                 continue;
706
707                         if (!buf->len) {
708                                 buf->ops = NULL;
709                                 ops->release(pipe, buf);
710                                 pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
711                                 pipe->nrbufs--;
712                                 if (pipe->inode)
713                                         do_wakeup = 1;
714                         }
715
716                         if (!sd->total_len)
717                                 break;
718                 }
719
720                 if (pipe->nrbufs)
721                         continue;
722                 if (!pipe->writers)
723                         break;
724                 if (!pipe->waiting_writers) {
725                         if (ret)
726                                 break;
727                 }
728
729                 if (sd->flags & SPLICE_F_NONBLOCK) {
730                         if (!ret)
731                                 ret = -EAGAIN;
732                         break;
733                 }
734
735                 if (signal_pending(current)) {
736                         if (!ret)
737                                 ret = -ERESTARTSYS;
738                         break;
739                 }
740
741                 if (do_wakeup) {
742                         smp_mb();
743                         if (waitqueue_active(&pipe->wait))
744                                 wake_up_interruptible_sync(&pipe->wait);
745                         kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
746                         do_wakeup = 0;
747                 }
748
749                 pipe_wait(pipe);
750         }
751
752         if (do_wakeup) {
753                 smp_mb();
754                 if (waitqueue_active(&pipe->wait))
755                         wake_up_interruptible(&pipe->wait);
756                 kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
757         }
758
759         return ret;
760 }
761 EXPORT_SYMBOL(__splice_from_pipe);
762
763 /**
764  * splice_from_pipe - splice data from a pipe to a file
765  * @pipe:       pipe to splice from
766  * @out:        file to splice to
767  * @ppos:       position in @out
768  * @len:        how many bytes to splice
769  * @flags:      splice modifier flags
770  * @actor:      handler that splices the data
771  *
772  * Description:
773  *    See __splice_from_pipe. This function locks the input and output inodes,
774  *    otherwise it's identical to __splice_from_pipe().
775  *
776  */
777 ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
778                          loff_t *ppos, size_t len, unsigned int flags,
779                          splice_actor *actor)
780 {
781         ssize_t ret;
782         struct inode *inode = out->f_mapping->host;
783         struct splice_desc sd = {
784                 .total_len = len,
785                 .flags = flags,
786                 .pos = *ppos,
787                 .u.file = out,
788         };
789
790         /*
791          * The actor worker might be calling ->prepare_write and
792          * ->commit_write. Most of the time, these expect i_mutex to
793          * be held. Since this may result in an ABBA deadlock with
794          * pipe->inode, we have to order lock acquiry here.
795          */
796         inode_double_lock(inode, pipe->inode);
797         ret = __splice_from_pipe(pipe, &sd, actor);
798         inode_double_unlock(inode, pipe->inode);
799
800         return ret;
801 }
802
803 /**
804  * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
805  * @pipe:       pipe info
806  * @out:        file to write to
807  * @ppos:       position in @out
808  * @len:        number of bytes to splice
809  * @flags:      splice modifier flags
810  *
811  * Description:
812  *    Will either move or copy pages (determined by @flags options) from
813  *    the given pipe inode to the given file. The caller is responsible
814  *    for acquiring i_mutex on both inodes.
815  *
816  */
817 ssize_t
818 generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
819                                  loff_t *ppos, size_t len, unsigned int flags)
820 {
821         struct address_space *mapping = out->f_mapping;
822         struct inode *inode = mapping->host;
823         struct splice_desc sd = {
824                 .total_len = len,
825                 .flags = flags,
826                 .pos = *ppos,
827                 .u.file = out,
828         };
829         ssize_t ret;
830         int err;
831
832         err = remove_suid(out->f_path.dentry);
833         if (unlikely(err))
834                 return err;
835
836         ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
837         if (ret > 0) {
838                 unsigned long nr_pages;
839
840                 *ppos += ret;
841                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
842
843                 /*
844                  * If file or inode is SYNC and we actually wrote some data,
845                  * sync it.
846                  */
847                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
848                         err = generic_osync_inode(inode, mapping,
849                                                   OSYNC_METADATA|OSYNC_DATA);
850
851                         if (err)
852                                 ret = err;
853                 }
854                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
855         }
856
857         return ret;
858 }
859
860 EXPORT_SYMBOL(generic_file_splice_write_nolock);
861
862 /**
863  * generic_file_splice_write - splice data from a pipe to a file
864  * @pipe:       pipe info
865  * @out:        file to write to
866  * @ppos:       position in @out
867  * @len:        number of bytes to splice
868  * @flags:      splice modifier flags
869  *
870  * Description:
871  *    Will either move or copy pages (determined by @flags options) from
872  *    the given pipe inode to the given file.
873  *
874  */
875 ssize_t
876 generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
877                           loff_t *ppos, size_t len, unsigned int flags)
878 {
879         struct address_space *mapping = out->f_mapping;
880         struct inode *inode = mapping->host;
881         ssize_t ret;
882         int err;
883
884         err = should_remove_suid(out->f_path.dentry);
885         if (unlikely(err)) {
886                 mutex_lock(&inode->i_mutex);
887                 err = __remove_suid(out->f_path.dentry, err);
888                 mutex_unlock(&inode->i_mutex);
889                 if (err)
890                         return err;
891         }
892
893         ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
894         if (ret > 0) {
895                 unsigned long nr_pages;
896
897                 *ppos += ret;
898                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
899
900                 /*
901                  * If file or inode is SYNC and we actually wrote some data,
902                  * sync it.
903                  */
904                 if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
905                         mutex_lock(&inode->i_mutex);
906                         err = generic_osync_inode(inode, mapping,
907                                                   OSYNC_METADATA|OSYNC_DATA);
908                         mutex_unlock(&inode->i_mutex);
909
910                         if (err)
911                                 ret = err;
912                 }
913                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
914         }
915
916         return ret;
917 }
918
919 EXPORT_SYMBOL(generic_file_splice_write);
920
921 /**
922  * generic_splice_sendpage - splice data from a pipe to a socket
923  * @pipe:       pipe to splice from
924  * @out:        socket to write to
925  * @ppos:       position in @out
926  * @len:        number of bytes to splice
927  * @flags:      splice modifier flags
928  *
929  * Description:
930  *    Will send @len bytes from the pipe to a network socket. No data copying
931  *    is involved.
932  *
933  */
934 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
935                                 loff_t *ppos, size_t len, unsigned int flags)
936 {
937         return splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_sendpage);
938 }
939
940 EXPORT_SYMBOL(generic_splice_sendpage);
941
942 /*
943  * Attempt to initiate a splice from pipe to file.
944  */
945 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
946                            loff_t *ppos, size_t len, unsigned int flags)
947 {
948         int ret;
949
950         if (unlikely(!out->f_op || !out->f_op->splice_write))
951                 return -EINVAL;
952
953         if (unlikely(!(out->f_mode & FMODE_WRITE)))
954                 return -EBADF;
955
956         ret = rw_verify_area(WRITE, out, ppos, len);
957         if (unlikely(ret < 0))
958                 return ret;
959
960         ret = security_file_permission(out, MAY_WRITE);
961         if (unlikely(ret < 0))
962                 return ret;
963
964         return out->f_op->splice_write(pipe, out, ppos, len, flags);
965 }
966
967 /*
968  * Attempt to initiate a splice from a file to a pipe.
969  */
970 static long do_splice_to(struct file *in, loff_t *ppos,
971                          struct pipe_inode_info *pipe, size_t len,
972                          unsigned int flags)
973 {
974         int ret;
975
976         if (unlikely(!in->f_op || !in->f_op->splice_read))
977                 return -EINVAL;
978
979         if (unlikely(!(in->f_mode & FMODE_READ)))
980                 return -EBADF;
981
982         ret = rw_verify_area(READ, in, ppos, len);
983         if (unlikely(ret < 0))
984                 return ret;
985
986         ret = security_file_permission(in, MAY_READ);
987         if (unlikely(ret < 0))
988                 return ret;
989
990         return in->f_op->splice_read(in, ppos, pipe, len, flags);
991 }
992
993 /**
994  * splice_direct_to_actor - splices data directly between two non-pipes
995  * @in:         file to splice from
996  * @sd:         actor information on where to splice to
997  * @actor:      handles the data splicing
998  *
999  * Description:
1000  *    This is a special case helper to splice directly between two
1001  *    points, without requiring an explicit pipe. Internally an allocated
1002  *    pipe is cached in the process, and reused during the life time of
1003  *    that process.
1004  *
1005  */
1006 ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
1007                                splice_direct_actor *actor)
1008 {
1009         struct pipe_inode_info *pipe;
1010         long ret, bytes;
1011         umode_t i_mode;
1012         size_t len;
1013         int i, flags;
1014
1015         /*
1016          * We require the input being a regular file, as we don't want to
1017          * randomly drop data for eg socket -> socket splicing. Use the
1018          * piped splicing for that!
1019          */
1020         i_mode = in->f_path.dentry->d_inode->i_mode;
1021         if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
1022                 return -EINVAL;
1023
1024         /*
1025          * neither in nor out is a pipe, setup an internal pipe attached to
1026          * 'out' and transfer the wanted data from 'in' to 'out' through that
1027          */
1028         pipe = current->splice_pipe;
1029         if (unlikely(!pipe)) {
1030                 pipe = alloc_pipe_info(NULL);
1031                 if (!pipe)
1032                         return -ENOMEM;
1033
1034                 /*
1035                  * We don't have an immediate reader, but we'll read the stuff
1036                  * out of the pipe right after the splice_to_pipe(). So set
1037                  * PIPE_READERS appropriately.
1038                  */
1039                 pipe->readers = 1;
1040
1041                 current->splice_pipe = pipe;
1042         }
1043
1044         /*
1045          * Do the splice.
1046          */
1047         ret = 0;
1048         bytes = 0;
1049         len = sd->total_len;
1050         flags = sd->flags;
1051
1052         /*
1053          * Don't block on output, we have to drain the direct pipe.
1054          */
1055         sd->flags &= ~SPLICE_F_NONBLOCK;
1056
1057         while (len) {
1058                 size_t read_len;
1059                 loff_t pos = sd->pos;
1060
1061                 ret = do_splice_to(in, &pos, pipe, len, flags);
1062                 if (unlikely(ret <= 0))
1063                         goto out_release;
1064
1065                 read_len = ret;
1066                 sd->total_len = read_len;
1067
1068                 /*
1069                  * NOTE: nonblocking mode only applies to the input. We
1070                  * must not do the output in nonblocking mode as then we
1071                  * could get stuck data in the internal pipe:
1072                  */
1073                 ret = actor(pipe, sd);
1074                 if (unlikely(ret <= 0))
1075                         goto out_release;
1076
1077                 bytes += ret;
1078                 len -= ret;
1079                 sd->pos = pos;
1080
1081                 if (ret < read_len)
1082                         goto out_release;
1083         }
1084
1085         pipe->nrbufs = pipe->curbuf = 0;
1086         return bytes;
1087
1088 out_release:
1089         /*
1090          * If we did an incomplete transfer we must release
1091          * the pipe buffers in question:
1092          */
1093         for (i = 0; i < PIPE_BUFFERS; i++) {
1094                 struct pipe_buffer *buf = pipe->bufs + i;
1095
1096                 if (buf->ops) {
1097                         buf->ops->release(pipe, buf);
1098                         buf->ops = NULL;
1099                 }
1100         }
1101         pipe->nrbufs = pipe->curbuf = 0;
1102
1103         /*
1104          * If we transferred some data, return the number of bytes:
1105          */
1106         if (bytes > 0)
1107                 return bytes;
1108
1109         return ret;
1110
1111 }
1112 EXPORT_SYMBOL(splice_direct_to_actor);
1113
1114 static int direct_splice_actor(struct pipe_inode_info *pipe,
1115                                struct splice_desc *sd)
1116 {
1117         struct file *file = sd->u.file;
1118
1119         return do_splice_from(pipe, file, &sd->pos, sd->total_len, sd->flags);
1120 }
1121
1122 /**
1123  * do_splice_direct - splices data directly between two files
1124  * @in:         file to splice from
1125  * @ppos:       input file offset
1126  * @out:        file to splice to
1127  * @len:        number of bytes to splice
1128  * @flags:      splice modifier flags
1129  *
1130  * Description:
1131  *    For use by do_sendfile(). splice can easily emulate sendfile, but
1132  *    doing it in the application would incur an extra system call
1133  *    (splice in + splice out, as compared to just sendfile()). So this helper
1134  *    can splice directly through a process-private pipe.
1135  *
1136  */
1137 long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
1138                       size_t len, unsigned int flags)
1139 {
1140         struct splice_desc sd = {
1141                 .len            = len,
1142                 .total_len      = len,
1143                 .flags          = flags,
1144                 .pos            = *ppos,
1145                 .u.file         = out,
1146         };
1147         long ret;
1148
1149         ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
1150         if (ret > 0)
1151                 *ppos += ret;
1152
1153         return ret;
1154 }
1155
1156 /*
1157  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
1158  * location, so checking ->i_pipe is not enough to verify that this is a
1159  * pipe.
1160  */
1161 static inline struct pipe_inode_info *pipe_info(struct inode *inode)
1162 {
1163         if (S_ISFIFO(inode->i_mode))
1164                 return inode->i_pipe;
1165
1166         return NULL;
1167 }
1168
1169 /*
1170  * Determine where to splice to/from.
1171  */
1172 static long do_splice(struct file *in, loff_t __user *off_in,
1173                       struct file *out, loff_t __user *off_out,
1174                       size_t len, unsigned int flags)
1175 {
1176         struct pipe_inode_info *pipe;
1177         loff_t offset, *off;
1178         long ret;
1179
1180         pipe = pipe_info(in->f_path.dentry->d_inode);
1181         if (pipe) {
1182                 if (off_in)
1183                         return -ESPIPE;
1184                 if (off_out) {
1185                         if (out->f_op->llseek == no_llseek)
1186                                 return -EINVAL;
1187                         if (copy_from_user(&offset, off_out, sizeof(loff_t)))
1188                                 return -EFAULT;
1189                         off = &offset;
1190                 } else
1191                         off = &out->f_pos;
1192
1193                 ret = do_splice_from(pipe, out, off, len, flags);
1194
1195                 if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
1196                         ret = -EFAULT;
1197
1198                 return ret;
1199         }
1200
1201         pipe = pipe_info(out->f_path.dentry->d_inode);
1202         if (pipe) {
1203                 if (off_out)
1204                         return -ESPIPE;
1205                 if (off_in) {
1206                         if (in->f_op->llseek == no_llseek)
1207                                 return -EINVAL;
1208                         if (copy_from_user(&offset, off_in, sizeof(loff_t)))
1209                                 return -EFAULT;
1210                         off = &offset;
1211                 } else
1212                         off = &in->f_pos;
1213
1214                 ret = do_splice_to(in, off, pipe, len, flags);
1215
1216                 if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
1217                         ret = -EFAULT;
1218
1219                 return ret;
1220         }
1221
1222         return -EINVAL;
1223 }
1224
1225 /*
1226  * Map an iov into an array of pages and offset/length tupples. With the
1227  * partial_page structure, we can map several non-contiguous ranges into
1228  * our ones pages[] map instead of splitting that operation into pieces.
1229  * Could easily be exported as a generic helper for other users, in which
1230  * case one would probably want to add a 'max_nr_pages' parameter as well.
1231  */
1232 static int get_iovec_page_array(const struct iovec __user *iov,
1233                                 unsigned int nr_vecs, struct page **pages,
1234                                 struct partial_page *partial, int aligned)
1235 {
1236         int buffers = 0, error = 0;
1237
1238         /*
1239          * It's ok to take the mmap_sem for reading, even
1240          * across a "get_user()".
1241          */
1242         down_read(&current->mm->mmap_sem);
1243
1244         while (nr_vecs) {
1245                 unsigned long off, npages;
1246                 void __user *base;
1247                 size_t len;
1248                 int i;
1249
1250                 /*
1251                  * Get user address base and length for this iovec.
1252                  */
1253                 error = get_user(base, &iov->iov_base);
1254                 if (unlikely(error))
1255                         break;
1256                 error = get_user(len, &iov->iov_len);
1257                 if (unlikely(error))
1258                         break;
1259
1260                 /*
1261                  * Sanity check this iovec. 0 read succeeds.
1262                  */
1263                 if (unlikely(!len))
1264                         break;
1265                 error = -EFAULT;
1266                 if (unlikely(!base))
1267                         break;
1268
1269                 /*
1270                  * Get this base offset and number of pages, then map
1271                  * in the user pages.
1272                  */
1273                 off = (unsigned long) base & ~PAGE_MASK;
1274
1275                 /*
1276                  * If asked for alignment, the offset must be zero and the
1277                  * length a multiple of the PAGE_SIZE.
1278                  */
1279                 error = -EINVAL;
1280                 if (aligned && (off || len & ~PAGE_MASK))
1281                         break;
1282
1283                 npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1284                 if (npages > PIPE_BUFFERS - buffers)
1285                         npages = PIPE_BUFFERS - buffers;
1286
1287                 error = get_user_pages(current, current->mm,
1288                                        (unsigned long) base, npages, 0, 0,
1289                                        &pages[buffers], NULL);
1290
1291                 if (unlikely(error <= 0))
1292                         break;
1293
1294                 /*
1295                  * Fill this contiguous range into the partial page map.
1296                  */
1297                 for (i = 0; i < error; i++) {
1298                         const int plen = min_t(size_t, len, PAGE_SIZE - off);
1299
1300                         partial[buffers].offset = off;
1301                         partial[buffers].len = plen;
1302
1303                         off = 0;
1304                         len -= plen;
1305                         buffers++;
1306                 }
1307
1308                 /*
1309                  * We didn't complete this iov, stop here since it probably
1310                  * means we have to move some of this into a pipe to
1311                  * be able to continue.
1312                  */
1313                 if (len)
1314                         break;
1315
1316                 /*
1317                  * Don't continue if we mapped fewer pages than we asked for,
1318                  * or if we mapped the max number of pages that we have
1319                  * room for.
1320                  */
1321                 if (error < npages || buffers == PIPE_BUFFERS)
1322                         break;
1323
1324                 nr_vecs--;
1325                 iov++;
1326         }
1327
1328         up_read(&current->mm->mmap_sem);
1329
1330         if (buffers)
1331                 return buffers;
1332
1333         return error;
1334 }
1335
1336 static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
1337                         struct splice_desc *sd)
1338 {
1339         char *src;
1340         int ret;
1341
1342         ret = buf->ops->confirm(pipe, buf);
1343         if (unlikely(ret))
1344                 return ret;
1345
1346         /*
1347          * See if we can use the atomic maps, by prefaulting in the
1348          * pages and doing an atomic copy
1349          */
1350         if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
1351                 src = buf->ops->map(pipe, buf, 1);
1352                 ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
1353                                                         sd->len);
1354                 buf->ops->unmap(pipe, buf, src);
1355                 if (!ret) {
1356                         ret = sd->len;
1357                         goto out;
1358                 }
1359         }
1360
1361         /*
1362          * No dice, use slow non-atomic map and copy
1363          */
1364         src = buf->ops->map(pipe, buf, 0);
1365
1366         ret = sd->len;
1367         if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
1368                 ret = -EFAULT;
1369
1370 out:
1371         if (ret > 0)
1372                 sd->u.userptr += ret;
1373         buf->ops->unmap(pipe, buf, src);
1374         return ret;
1375 }
1376
1377 /*
1378  * For lack of a better implementation, implement vmsplice() to userspace
1379  * as a simple copy of the pipes pages to the user iov.
1380  */
1381 static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
1382                              unsigned long nr_segs, unsigned int flags)
1383 {
1384         struct pipe_inode_info *pipe;
1385         struct splice_desc sd;
1386         ssize_t size;
1387         int error;
1388         long ret;
1389
1390         pipe = pipe_info(file->f_path.dentry->d_inode);
1391         if (!pipe)
1392                 return -EBADF;
1393
1394         if (pipe->inode)
1395                 mutex_lock(&pipe->inode->i_mutex);
1396
1397         error = ret = 0;
1398         while (nr_segs) {
1399                 void __user *base;
1400                 size_t len;
1401
1402                 /*
1403                  * Get user address base and length for this iovec.
1404                  */
1405                 error = get_user(base, &iov->iov_base);
1406                 if (unlikely(error))
1407                         break;
1408                 error = get_user(len, &iov->iov_len);
1409                 if (unlikely(error))
1410                         break;
1411
1412                 /*
1413                  * Sanity check this iovec. 0 read succeeds.
1414                  */
1415                 if (unlikely(!len))
1416                         break;
1417                 if (unlikely(!base)) {
1418                         error = -EFAULT;
1419                         break;
1420                 }
1421
1422                 sd.len = 0;
1423                 sd.total_len = len;
1424                 sd.flags = flags;
1425                 sd.u.userptr = base;
1426                 sd.pos = 0;
1427
1428                 size = __splice_from_pipe(pipe, &sd, pipe_to_user);
1429                 if (size < 0) {
1430                         if (!ret)
1431                                 ret = size;
1432
1433                         break;
1434                 }
1435
1436                 ret += size;
1437
1438                 if (size < len)
1439                         break;
1440
1441                 nr_segs--;
1442                 iov++;
1443         }
1444
1445         if (pipe->inode)
1446                 mutex_unlock(&pipe->inode->i_mutex);
1447
1448         if (!ret)
1449                 ret = error;
1450
1451         return ret;
1452 }
1453
1454 /*
1455  * vmsplice splices a user address range into a pipe. It can be thought of
1456  * as splice-from-memory, where the regular splice is splice-from-file (or
1457  * to file). In both cases the output is a pipe, naturally.
1458  */
1459 static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
1460                              unsigned long nr_segs, unsigned int flags)
1461 {
1462         struct pipe_inode_info *pipe;
1463         struct page *pages[PIPE_BUFFERS];
1464         struct partial_page partial[PIPE_BUFFERS];
1465         struct splice_pipe_desc spd = {
1466                 .pages = pages,
1467                 .partial = partial,
1468                 .flags = flags,
1469                 .ops = &user_page_pipe_buf_ops,
1470         };
1471
1472         pipe = pipe_info(file->f_path.dentry->d_inode);
1473         if (!pipe)
1474                 return -EBADF;
1475
1476         spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
1477                                             flags & SPLICE_F_GIFT);
1478         if (spd.nr_pages <= 0)
1479                 return spd.nr_pages;
1480
1481         return splice_to_pipe(pipe, &spd);
1482 }
1483
1484 /*
1485  * Note that vmsplice only really supports true splicing _from_ user memory
1486  * to a pipe, not the other way around. Splicing from user memory is a simple
1487  * operation that can be supported without any funky alignment restrictions
1488  * or nasty vm tricks. We simply map in the user memory and fill them into
1489  * a pipe. The reverse isn't quite as easy, though. There are two possible
1490  * solutions for that:
1491  *
1492  *      - memcpy() the data internally, at which point we might as well just
1493  *        do a regular read() on the buffer anyway.
1494  *      - Lots of nasty vm tricks, that are neither fast nor flexible (it
1495  *        has restriction limitations on both ends of the pipe).
1496  *
1497  * Currently we punt and implement it as a normal copy, see pipe_to_user().
1498  *
1499  */
1500 asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
1501                              unsigned long nr_segs, unsigned int flags)
1502 {
1503         struct file *file;
1504         long error;
1505         int fput;
1506
1507         if (unlikely(nr_segs > UIO_MAXIOV))
1508                 return -EINVAL;
1509         else if (unlikely(!nr_segs))
1510                 return 0;
1511
1512         error = -EBADF;
1513         file = fget_light(fd, &fput);
1514         if (file) {
1515                 if (file->f_mode & FMODE_WRITE)
1516                         error = vmsplice_to_pipe(file, iov, nr_segs, flags);
1517                 else if (file->f_mode & FMODE_READ)
1518                         error = vmsplice_to_user(file, iov, nr_segs, flags);
1519
1520                 fput_light(file, fput);
1521         }
1522
1523         return error;
1524 }
1525
1526 asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
1527                            int fd_out, loff_t __user *off_out,
1528                            size_t len, unsigned int flags)
1529 {
1530         long error;
1531         struct file *in, *out;
1532         int fput_in, fput_out;
1533
1534         if (unlikely(!len))
1535                 return 0;
1536
1537         error = -EBADF;
1538         in = fget_light(fd_in, &fput_in);
1539         if (in) {
1540                 if (in->f_mode & FMODE_READ) {
1541                         out = fget_light(fd_out, &fput_out);
1542                         if (out) {
1543                                 if (out->f_mode & FMODE_WRITE)
1544                                         error = do_splice(in, off_in,
1545                                                           out, off_out,
1546                                                           len, flags);
1547                                 fput_light(out, fput_out);
1548                         }
1549                 }
1550
1551                 fput_light(in, fput_in);
1552         }
1553
1554         return error;
1555 }
1556
1557 /*
1558  * Make sure there's data to read. Wait for input if we can, otherwise
1559  * return an appropriate error.
1560  */
1561 static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1562 {
1563         int ret;
1564
1565         /*
1566          * Check ->nrbufs without the inode lock first. This function
1567          * is speculative anyways, so missing one is ok.
1568          */
1569         if (pipe->nrbufs)
1570                 return 0;
1571
1572         ret = 0;
1573         mutex_lock(&pipe->inode->i_mutex);
1574
1575         while (!pipe->nrbufs) {
1576                 if (signal_pending(current)) {
1577                         ret = -ERESTARTSYS;
1578                         break;
1579                 }
1580                 if (!pipe->writers)
1581                         break;
1582                 if (!pipe->waiting_writers) {
1583                         if (flags & SPLICE_F_NONBLOCK) {
1584                                 ret = -EAGAIN;
1585                                 break;
1586                         }
1587                 }
1588                 pipe_wait(pipe);
1589         }
1590
1591         mutex_unlock(&pipe->inode->i_mutex);
1592         return ret;
1593 }
1594
1595 /*
1596  * Make sure there's writeable room. Wait for room if we can, otherwise
1597  * return an appropriate error.
1598  */
1599 static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
1600 {
1601         int ret;
1602
1603         /*
1604          * Check ->nrbufs without the inode lock first. This function
1605          * is speculative anyways, so missing one is ok.
1606          */
1607         if (pipe->nrbufs < PIPE_BUFFERS)
1608                 return 0;
1609
1610         ret = 0;
1611         mutex_lock(&pipe->inode->i_mutex);
1612
1613         while (pipe->nrbufs >= PIPE_BUFFERS) {
1614                 if (!pipe->readers) {
1615                         send_sig(SIGPIPE, current, 0);
1616                         ret = -EPIPE;
1617                         break;
1618                 }
1619                 if (flags & SPLICE_F_NONBLOCK) {
1620                         ret = -EAGAIN;
1621                         break;
1622                 }
1623                 if (signal_pending(current)) {
1624                         ret = -ERESTARTSYS;
1625                         break;
1626                 }
1627                 pipe->waiting_writers++;
1628                 pipe_wait(pipe);
1629                 pipe->waiting_writers--;
1630         }
1631
1632         mutex_unlock(&pipe->inode->i_mutex);
1633         return ret;
1634 }
1635
1636 /*
1637  * Link contents of ipipe to opipe.
1638  */
1639 static int link_pipe(struct pipe_inode_info *ipipe,
1640                      struct pipe_inode_info *opipe,
1641                      size_t len, unsigned int flags)
1642 {
1643         struct pipe_buffer *ibuf, *obuf;
1644         int ret = 0, i = 0, nbuf;
1645
1646         /*
1647          * Potential ABBA deadlock, work around it by ordering lock
1648          * grabbing by inode address. Otherwise two different processes
1649          * could deadlock (one doing tee from A -> B, the other from B -> A).
1650          */
1651         inode_double_lock(ipipe->inode, opipe->inode);
1652
1653         do {
1654                 if (!opipe->readers) {
1655                         send_sig(SIGPIPE, current, 0);
1656                         if (!ret)
1657                                 ret = -EPIPE;
1658                         break;
1659                 }
1660
1661                 /*
1662                  * If we have iterated all input buffers or ran out of
1663                  * output room, break.
1664                  */
1665                 if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
1666                         break;
1667
1668                 ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
1669                 nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
1670
1671                 /*
1672                  * Get a reference to this pipe buffer,
1673                  * so we can copy the contents over.
1674                  */
1675                 ibuf->ops->get(ipipe, ibuf);
1676
1677                 obuf = opipe->bufs + nbuf;
1678                 *obuf = *ibuf;
1679
1680                 /*
1681                  * Don't inherit the gift flag, we need to
1682                  * prevent multiple steals of this page.
1683                  */
1684                 obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
1685
1686                 if (obuf->len > len)
1687                         obuf->len = len;
1688
1689                 opipe->nrbufs++;
1690                 ret += obuf->len;
1691                 len -= obuf->len;
1692                 i++;
1693         } while (len);
1694
1695         inode_double_unlock(ipipe->inode, opipe->inode);
1696
1697         /*
1698          * If we put data in the output pipe, wakeup any potential readers.
1699          */
1700         if (ret > 0) {
1701                 smp_mb();
1702                 if (waitqueue_active(&opipe->wait))
1703                         wake_up_interruptible(&opipe->wait);
1704                 kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
1705         }
1706
1707         return ret;
1708 }
1709
1710 /*
1711  * This is a tee(1) implementation that works on pipes. It doesn't copy
1712  * any data, it simply references the 'in' pages on the 'out' pipe.
1713  * The 'flags' used are the SPLICE_F_* variants, currently the only
1714  * applicable one is SPLICE_F_NONBLOCK.
1715  */
1716 static long do_tee(struct file *in, struct file *out, size_t len,
1717                    unsigned int flags)
1718 {
1719         struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
1720         struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
1721         int ret = -EINVAL;
1722
1723         /*
1724          * Duplicate the contents of ipipe to opipe without actually
1725          * copying the data.
1726          */
1727         if (ipipe && opipe && ipipe != opipe) {
1728                 /*
1729                  * Keep going, unless we encounter an error. The ipipe/opipe
1730                  * ordering doesn't really matter.
1731                  */
1732                 ret = link_ipipe_prep(ipipe, flags);
1733                 if (!ret) {
1734                         ret = link_opipe_prep(opipe, flags);
1735                         if (!ret) {
1736                                 ret = link_pipe(ipipe, opipe, len, flags);
1737                                 if (!ret && (flags & SPLICE_F_NONBLOCK))
1738                                         ret = -EAGAIN;
1739                         }
1740                 }
1741         }
1742
1743         return ret;
1744 }
1745
1746 asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
1747 {
1748         struct file *in;
1749         int error, fput_in;
1750
1751         if (unlikely(!len))
1752                 return 0;
1753
1754         error = -EBADF;
1755         in = fget_light(fdin, &fput_in);
1756         if (in) {
1757                 if (in->f_mode & FMODE_READ) {
1758                         int fput_out;
1759                         struct file *out = fget_light(fdout, &fput_out);
1760
1761                         if (out) {
1762                                 if (out->f_mode & FMODE_WRITE)
1763                                         error = do_tee(in, out, len, flags);
1764                                 fput_light(out, fput_out);
1765                         }
1766                 }
1767                 fput_light(in, fput_in);
1768         }
1769
1770         return error;
1771 }